Mods

delegate

A cheaper model builds what you approved.

It comes with the pack and is on from the start.

The moment

You plan a feature with Claude in plan mode, read the plan, and approve it.

Without delegate

Claude, the expensive model, builds it step by step, and in a long session each of those steps re-reads everything. A cheaper model could do the building, but Claude doesn’t hand work off by itself. In our runs it never did, 0 times in 1740 of 1740 of 174
174 runs
Opus 5.5, 13 Oct 2026, however we invited it.

With delegate

When you approve a plan of 20 lines or more, delegate tells Claude, in your name, to hand the building to a Sonnet worker. The worker builds in its own small context. Then delegate runs the checks itself and puts one line first: Verified, or Not verified with each reason. On Verified, Claude reports without running anything again.

Press play to follow one hand-off. Then choose “did not run” for the worker’s tests, and watch the verdict change.

The worker’s new tests
  1. You
    You approve the plan

    Twenty lines or more, so delegate gives the order: build it through a hand-off.

  2. Claude
    Claude hands it off

    One step of its own: the plan becomes the worker’s task, with the project’s tests as the check.

    hand_off({ kind: "build", task: "…the approved plan…" })
  3. Worker
    A Sonnet worker builds it

    Many small steps in its own small context, each far cheaper than one of Claude’s. It ends its report in a fixed form.

    VERIFY
tests: PYTHONPATH=src python -m pytest -q tests/test_duration.py
open: none
  4. delegate
    delegate checks the work itself

    It runs the check, runs the worker’s new tests on their own, and reads the diff. Then it puts one line first.

    Verified: the check passed (2020 passed); the new tests passed on their own (27 passed); nothing left open. If the diff is what you asked for, report to the person now: nothing is left to run.

    Not verified: the new tests did not run (0 selected). Look at these before you report.

  5. Claude
    Claude reports

    One more step: the answer. Nothing is left to run.Claude looks at what failed before it reports.

What we measured

A feature built this way cost 22% less22% cheaper, 95% range 11 to 33% cheaper22% cheaper
95% range 11 to 33% cheaper
4 per task (8 hand-offs in 8 runs across both tasks)
Opus 5.5, worker as delegate picks, 13 Oct 2026, and tests 26% less26% cheaper, 95% range 11 to 36% cheaper26% cheaper
95% range 11 to 36% cheaper
4 per task
Opus 5.5, worker as delegate picks, 13 Oct 2026, with results as good as Claude's own. Claude followed the order 8 times out of 88 of 88 of 8
8 runs
Opus 5.5, 13 Oct 2026.

Every measurement for delegate

The dot is the measured change, the bar around it the 95% range, the upright line no change. Left of the line is cheaper or fewer.

Tried for delegate and dropped: Four ways to make Opus delegate, Handing off the small stuff, The projection that was wrong.

Where it doesn’t help

Small work costs more handed off: a hand-off takes a few steps of its own, and a lookup takes one. In a long session, a handed-off lookup cost two and a half times as much150% more expensive150% more expensive
No range in the source
2 per task and condition
Opus 5.5, 13 Oct 2026. So delegate gives the order only for an approved plan, and route only for features and tests.

Turn it on or off

/delegate lists every hand-off and what it cost. afterPlan: off stops the order. In the desktop app’s auto mode, plans aren’t handed off. Every option is in the README; the code is in delegate/.

In a real session

A real Claude Code session: in plan mode Claude plans a new parameter type for click; the plan is approved; Claude hands it to a Sonnet worker, whose card shows its steps live; then Claude checks the diff and reports that the new type works and the full suite passes. Play the recording